bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
1
2
Title:
Determining causal genes from GWAS signals using topologically associating domains
3
4
Authors:
5
Gregory P. Way1,2, Daniel W. Youngstrom3, Kurt D. Hankenson3, Casey S. Greene2*, and
6
Struan F.A. Grant4,5*
7
* These authors directed this work jointly
8
9
Affiliations:
10
1
11
Philadelphia, PA 19104, USA
12
2
13
Pennsylvania, Philadelphia, PA 19104, USA
14
3
15
State University, East Lansing, MI 48824, USA
16
4
17
Philadelphia, PA 19104, USA
18
5
19
Hospital of Philadelphia, Philadelphia, PA 19104, USA
Genomics and Computational Biology Graduate Program, University of Pennsylvania,
Department of Systems Pharmacology and Translational Therapeutics, University of
Department of Small Animal Clinical Sciences, College of Veterinary Medicine, Michigan
Department of Pediatrics, Perelman School of Medicine, University of Pennsylvania,
Division of Human Genetics, Division of Endocrinology and Diabetes, The Children's
20
21
Author Emails:
22
GPW: [email protected]; DWY: [email protected]; KDH: [email protected];
23
CSG: [email protected]; *SFAG: [email protected]
1
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
24
Co-Corresponding Authors:*
Struan F.A. Grant
Casey S. Greene
Rm 1102D, 3615 Civic Center Blvd
10-131 SCTR 34th and Civic Center Blvd,
Philadelphia, PA 19104
Philadelphia, PA 19104
Office: 267-426-2795
Office: 215-573-2991
Fax: 215-590-1258
Fax: 215-573-9135
25
26
Abstract:
27
Background
28
Genome wide association studies (GWAS) have contributed significantly to the field of
29
complex disease genetics. However, GWAS only report signals associated with a given trait and
30
do not necessarily identify the precise location of culprit genes. As most association signals
31
occur in non-coding regions of the genome, it is often challenging to assign genomic variants to
32
the underlying causal mechanism(s). Topologically associating domains (TADs) are primarily
33
cell-type independent genomic regions that define interactome boundaries and can aid in the
34
designation of limits within which a GWAS locus most likely impacts gene function.
35
Results
36
We describe and validate a computational method that uses the genic content of TADs to
37
assign GWAS signals to likely causal genes. Our method, called “TAD_Pathways”, performs a
38
Gene Ontology (GO) analysis over all genes that reside within the boundaries of all TADs
39
corresponding to the GWAS signals for a given trait or disease. We applied our pipeline to the
40
GWAS catalog entries associated with bone mineral density (BMD), identifying ‘Skeletal
41
System Development’ (Benjamini-Hochberg adjusted p = 1.02x10-5) as the top ranked pathway.
42
Often, the causal gene identified at a given locus was well known and/or the nearest gene to the
43
sentinel SNP. In other cases, our method implicated a gene further away. Our molecular
2
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
44
experiments describe a novel example: ACP2, implicated at the canonical ‘ARHGAP1’ locus. We
45
found ACP2 to be an important regulator of osteoblast metabolism, whereas a causal role of
46
ARHGAP1 was not supported.
47
Conclusions
48
Our results demonstrate how basic principles of three-dimensional genome organization can
49
help define biologically informed windows of signal association. We anticipate that
50
incorporating TADs will aid in refining and improving the performance of a variety of
51
algorithms that linearly interpret genomic content.
52
53
Keywords:
54
Genome wide association study, gene prioritization, topologically associating domains,
55
pathway analysis, bone mineral density
56
57
Background:
58
Genome-wide association studies (GWAS) have been applied to over 300 different traits,
59
leading to the discovery and subsequent validation of several important disease associations [1].
60
However, GWAS can only discover association signals in the data. Subsequent assignment of
61
signal to causal genes has proven difficult due to these signals falling principally within
62
noncoding genomic regions [2–4] and not necessarily implicating the nearest gene [5]. For
63
example, a signal found within an intron for FTO, a well-studied gene previously thought to be
64
important for obesity [6], has been shown to physically interact with and lead to the differential
65
expression of two genes (IRX3 and IRX5) directly next to this gene, and not FTO itself [7–9].
66
Moreover, there is evidence suggesting a type 2 diabetes GWAS association previously
67
implicating TCF7L2 [10] also influences the nearby ACSL5 gene [11]. It remains unclear how
3
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
68
pervasive these kinds of associations are, but similar strategies are necessary in order for GWAS
69
to better guide research and precision medicine [12].
70
Three-dimensional genomics has changed the way geneticists think about genome
71
organization and its functional implications [13,14]. Genome-wide chromatin interaction maps
72
have facilitated the development of several genome organization principles, including
73
topologically associating domains (TADs) [15–18]. TADs are sub-architectural units of the
74
overall genome organization that have consistent and functionally important genomic element
75
distributions including an enrichment of housekeeping genes, insulator elements, and early
76
replication timing regions at boundary regions [19–21]. TADs are largely consistent across
77
different cell types and demonstrate synteny [22,23]. These observations can therefore allow the
78
leveraging of TADs to set the bounds of where non-coding causal variants can most likely
79
impact promoters, enhancers and genes in a tissue independent fashion [24,25]. Therefore, we
80
sought to develop a method that integrates GWAS data with interactome boundaries to more
81
accurately map signals to the mostly likely candidate gene(s).
82
We developed a computational approach, called “TAD_Pathways”, which is agnostic to gene
83
locations relative to each GWAS signal within TADs. We scanned publically available GWAS
84
data for given traits and used TAD boundaries to output lists of genes likely to be causal. We
85
demonstrate this approach by assessing the influence of GWAS signals on bone mineral density
86
(BMD) [26–29]. This trait is clinically of great importance as low BMD is an important
87
precursor to osteoporosis, a disease condition affecting millions of patients annually [30]. We
88
also chose BMD as a trait for analysis because BMD GWAS primarily points to very well-
89
known genes involved in bone development (positive controls) but there remain a number of
90
established loci where no obvious gene resides, therefore offering the opportunity to uncover
4
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
91
novel biology. After applying our TAD_Pathways discovery approach, we investigated putative
92
causal genes using cell culture-based assays, identifying ACP2 as a novel regulator of osteoblast
93
metabolism.
94
95
Results:
96
The Genomic Landscape of SNPs across Topologically Associating Domains
97
We observed a consistent and non-random distribution of SNPs across TADs derived from
98
human embryonic stem cells (hESC), human fibroblasts (IMR90), mouse embryonic stem cells
99
(mESC), and mouse cortex cells (mcortex) cells. As expected, SNPs are tightly associated with
100
TAD length for each cell type, but there are substantial outlier TADs (Figure 1). For example,
101
the TAD harboring the largest number of common SNPs (minor allele frequency (MAF) greater
102
than 0.05) in hESC is located on chromosome 6 (UCSC hg19: chr6:31492021-32932022) and
103
has 19,431 SNPs. Not surprisingly, this TAD harbors an abundance of genes including HLA
104
genes, which are well known to have many polymorphic sites [31]. However, the other human
105
cell line (IMR90) outlier TAD is located on chromosome 8 (UCSC hg19: chr8:2132593-
106
6252592) and has 27,220 SNPs and could be potentially biologically meaningful. Indeed,
107
although this TAD harbors relatively few genes, it does include CSMD1, a gene implicated in
108
cancer and neurological disorders such as epilepsy and schizophrenia [32,33].
109
Common SNPs were enriched near the center of TADs (Figure 1A). This is the opposite of
110
gene (Supplementary Figure S1) and repeat element (Supplementary Figure S2) distributions
111
(also see Dixon et al. 2012) [22]. The repeat element distribution was driven largely by the
112
SINE/Alu repeat distribution, which could not be explained by GC content and estimated
113
evolutionary divergence (Supplementary Figure S3). We also observed that common SNPs are
5
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
114
significantly enriched in the 3’ half of TADs in hESC and mcortex cells (Supplementary Table
115
S1). There was also a slight increase in GWAS implicated SNPs near hESC TAD boundaries
116
(Figure 1B). Given the non-random patterns observed across the TADs, we went on to explore
117
the gene content further in an attempt to imply causality at given GWAS loci.
118
119
120
121
122
123
124
125
126
127
128
129
Figure 1. Distributions of single nucleotide polymorphisms (SNPs) across topologically
associating domains of four cell types. (A) Top – The length of TADs is associated with the
number of SNPs found in each cell type. Bars on the top and right side of the plots represent
histograms of each respective metric. hESC and IMR90 TADs are based on 1000 genomes phase
3 SNPs (hg19) and mESC and mcortex TADs represent 15 strains from the mouse genomes
project version 2 (mm9). Bottom – Number of SNPs found in each cell type. (B) Number of
GWAS SNPs from the NHGRI-EBI GWAS catalog that reached genome-wide significance in
replication required journals. In each case, we independently discretized TADs into 50 bins
where the distribution of elements is linear from 5’ to 3’ (bin 0 is the 5’ most end of all TADs).
130
6
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
131
TAD_Pathways reveals potentially causal genes within phenotype-associated TADs
132
Seeking to leverage TADs and disease associated SNPs, we integrated GWAS and TAD
133
domain boundaries in an effort to assign GWAS signals to causal genes. Alternative approaches
134
to understand the gene landscape of a locus that do not consider TAD boundaries typically either
135
assign genes to a GWAS signal based on nearest gene [34] or by an arbitrary or a linkage
136
disequilibrium-based window of several kilobases [35,36] (Figure 2A). Instead, we used TAD
137
boundaries and the full catalog of GWAS findings for a given trait or disease to assign genes to
138
GWAS variants based on overrepresentation in a gene set in an approach we termed
139
“TAD_Pathways”. For a given trait or disease, we collected all genes that are located in TADs
140
harboring significant GWAS signals. We then applied a statistical enrichment analysis for
141
biological pathways using this TAD gene set and assign candidate genes within a TAD based on
142
the pathways significantly associated with a phenotype (Figure 2B). In our implementation, we
143
used GO biological processes, GO cellular components, and GO molecular functions to provide
144
the pathway sets [37]. We included both experimentally confirmed and computationally inferred
145
GO gene annotations, which permit the inclusion of putative casual genes that do not necessarily
146
have literature support but are predicted by a variety of computational methods.
147
To validate our approach, we applied TAD_Pathways to bone mineral density (BMD)
148
GWAS results derived from replication-requiring journals [26–29]. Our method implicated
149
‘Skeletal System Development’ (Benjamini-Hochberg adjusted p = 1.02x10-5) as the top ranked
150
pathway. We provide full TAD_Pathways results for BMD in Supplementary Table S2.
151
Despite a high content of presumably non-causal genes, which we expect would contribute noise
152
to the overrepresentation analysis [38], our method demonstrated enrichment of a skeletal system
153
related pathway and selected a subset of potentially causal genes belonging to the same pathway.
7
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
154
Many of these genes (24/38) were not the nearest gene to the GWAS signal and several also had
155
independent expression quantitative trail loci (eQTL) support (Supplementary Table S3,
156
Supplementary Figure S4).
157
158
159
160
161
162
163
164
165
166
167
168
169
Figure 2. Concepts motivating our approach. Topologically associating domains (TADS) are
shown as orange triangles, genes are shown as black lines, and a genome wide significant
GWAS signal is shown as a dotted red line. (A) Three hypothetical examples illustrated by a
cartoon. The ground truth causal gene is shaded in red. The method-specific selected genes are
shaded in blue. The top panel describes a nearest gene approach. The nearest gene in this
scenario is not the gene actually impacted by the GWAS SNP. The middle panel describes a
window approach. Based either on linkage disequilibrium or an arbitrarily sized window, the
scenario does not capture the true gene. The bottom panel describes the TAD_Pathways
approach. In this scenario, the causal gene is selected for downstream assessment. (B) The
TAD_Pathways method. An example using Bone Mineral Density GWAS signals is shown.
170
siRNA Knockdown of TAD Pathway Gene Predictions in Osteoblast Cells
171
The loci rs7932354 (cytoband: 11p11.2) and rs11602954 (cytoband: 11p15.5) are currently
172
assigned to ARHGAP1 and BETL1 but our method implicated ACP2 and DEAF1, respectively.
173
The two genes implicated by TAD_Pathways, ACP2 and DEAF1, lacked eQTL support and were
174
not the nearest gene to the BMD GWAS signal. We tested the gene expression activity and
175
metabolic importance of these four genes, ARHGAP1, BETL1, DEAF1, and ACP2. Specifically,
176
our assays in a human fetal osteoblast cell line (hFOB) evaluate whether or not the
8
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
177
TAD_Pathways method identifies causal genes at GWAS signals beyond those captured by
178
closest and eQTL connected genes. Though these two genes were annotated to the identified GO
179
process, the annotation had been made computationally and their known biology did not provide
180
obvious links to bone biology.
181
We targeted the expression of all four of these genes in vitro using small interfering RNA
182
(siRNA), assessing knockdown efficiency at the mRNA level relative to untreated controls and
183
determined corresponding p values relative to scrambled siRNA controls. We used an siRNA
184
targeting tissue-nonspecific alkaline phosphatase (TNAP) as a positive control. Knockdown
185
efficiencies were: TNAP siRNA 48.7±9.9% (p=0.141), ARHGAP1 siRNA 68.7±14.3%
186
(p=0.015), ACP2 siRNA 48.9±6.4% (p=0.035), BET1L 56.4±1.0% (N.S.) and DEAF1
187
52.7±9.2% (p=0.021) (Figure 3). siRNA targeted against each gene of interest did not down-
188
regulate the expression of the other genes under investigation, indicating specificity of
189
knockdown, although we noted that TNAP siRNA did reduce DEAF1 gene expression, though
190
this did not reach the threshold for statistical significance (p= 0.077).
191
We noted significant variation across the three controls, with the scrambled siRNA control
192
altering expression of OCN (osteocalcin), IBSP (bone sialoprotein), TNAP and BET1L (p < 0.05).
193
Relative to the scrambled siRNA control, OCN was downregulated in all siRNA groups (p <
194
0.05) except for BET1L siRNA (p = 0.122). OSX, IBSP and TNAP were not significantly altered
195
by any siRNA treatment (Figure 3).
196
197
198
199
Metabolic Activity of TAD Pathway Gene Predictions
Use of ACP2 siRNA led to a 66.0% reduction in MTT metabolic activity versus the
scrambled siRNA control (p = 0.012). ARHGAP1 siRNA caused a 38.8% reduction, which fell
9
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
200
short of statistical significance (p = 0.088). siRNA targeted against TNAP, BET1L or DEAF1 did
201
not alter MTT metabolic activity (Figure 4A).
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
Figure 3: Real-time PCR of osteoblast differentiation genes and GWAS/TAD hits in hFOB cells.
siRNA was used to knock down expression of TNAP (positive control), ARHGAP1, ACP2,
BET1L and DEAF1. Relative expression of the osteoblast marker genes OSX, OCN and IBSP
suggest that GWAS/TAD hits are not major regulators of bone differentiation in this model. Red
bars highlight specificity of each siRNA knockdown. Values represent mean ± standard
deviation. Statistical significance relative to the scrambled siRNA control is annotated as: *p ≤
0.05 and #p ≤ 0.10 using a two-tailed Student’s t-test.
Figure 4. Validating two ‘TAD Pathway’ predictions for Bone Mineral Density GWAS hits on
hFOB cells. siRNA was used to knock down expression of TNAP, ARHGAP1, ACP2, BET1L and
DEAF1. (A) Knockdown of ACP2 decreases cellular metabolic activity, demonstrated using an
MTT assay. (B) ALP staining and quantitation indicates that knockdown of TNAP or ACP2
inhibits performance in an osteoblast differentiation assay. Values represent mean ± standard
deviation. Statistical significance relative to the scrambled siRNA control is annotated as: *p ≤
0.05 and #p ≤ 0.10 using a two-tailed Student’s t-test.
10
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
221
222
Influence of TAD_Pathways Gene Predictions on Alkaline Phosphatase Activity
Alkaline phosphatase (ALP) is highly expressed in osteoblasts; disruption of proliferation or
223
osteoblast differentiation would result in downregulation of ALP. Treatment with siRNA
224
resulted in changes in ALP staining that we analyzed further by quantitation. TNAP siRNA
225
significantly reduced ALP by 5.98±1.77 versus the scrambled siRNA control (p = 0.006). ACP2
226
siRNA also significantly reduced ALP intensity by 8.74±2.11 versus the scrambled siRNA
227
control (p = 0.003). The scrambled siRNA group stained less intensely than untreated or
228
transfection reagent control wells, but this did not reach statistical significance (0.05 < p < 0.10)
229
(Figure 4B).
230
231
232
Discussion:
We observed a nonrandom enrichment of SNPs in the center of TADs that was consistent
233
across different cell types, but was in the opposite direction of the gene and repeat elements
234
distributions. It is possible that the gene distribution is driving this phenomenon, since coding
235
regions are under higher evolutionary constraint and are thus more averse to SNPs [39].
236
Nevertheless, GWAS SNPs also appeared to be distributed closely to boundary regions in hESC
237
cells. This may support GWAS causally implicating nearest genes more frequently since genes
238
are also distributed near boundaries. The observation may also suggest that polymorphism in
239
regions near TAD boundaries are more important drivers of disease risk associations than
240
polymorphism in the center of TADs. However, we do not observe this pattern in IMR90 cells.
241
The SNP distribution was also opposite of the SINE/Alu repeat distribution. Given that Alu
242
elements tend to insert into GC-rich regions [40], we tracked GC content across TADs and
243
observed only a slight increase in GC content near TAD boundaries. There was also a slightly
11
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
244
inverse distribution of Alu evolutionary divergence [40]. Our results suggest that the Alu
245
distribution is primarily driven by intronic clustering [41] rather than GC-biased insertion or
246
evolutionary divergence. Recently, retrotransposons have been shown to act as genomic
247
insulators [42], while Alu repeats have been shown to be correlated with functional elements
248
[43]. However, the relationships between polymorphic sites, repeat elements, and genes across
249
TADs and higher level genome organization have yet to be explored in detail and warrants
250
further investigation.
251
TAD boundaries offer a unique computational opportunity to use biologically informed
252
windows to predefine areas of the genome that are more likely to interact with themselves. We
253
showed, as a proof of concept, that TADs can reveal functional GWAS variant to gene
254
relationships using BMD. Several of the TAD_Pathways implicated genes, including LRP5 and
255
other Wnt signaling genes, are bona fide BMD genes already identified by nearest gene GWAS,
256
eQTL analyses and human clinical syndromes [44,45], thus providing positive controls for our
257
approach. However, several BMD GWAS signals do not have obvious nearest gene associations,
258
which allowed us to validate our approach with two candidate causal, non-nearest gene
259
predictions: ACP2 and DEAF1. Both genes also did not have eQTL associations, but this is
260
likely a result of the eQTL browser lacking bone tissue.
261
To assess the validity of our predictions, we experimentally knocked down ACP2 and
262
DEAF1 in hFOB cells. siRNA for ACP2 and DEAF1 did not significantly alter expression of the
263
osteoblast marker genes OSX, IBSP or TNAP. OCN was downregulated in each of the
264
experimental groups relative to the scrambled siRNA control, but comparison with the reagent
265
control indicated no significant difference in any group, suggesting an off-target effect on OCN
266
in the scrambled siRNA group. Because osteoblast differentiation genes were not downregulated
12
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
267
following knockdown of the genes of interest, we concluded that these genes do not directly
268
regulate the transcriptional processes of osteoblast differentiation in vitro. The decrease in
269
DEAF1 expression following TNAP siRNA treatment, though not statistically significant,
270
suggests that DEAF1 may function downstream of TNAP.
271
There was a pronounced and statistically significant reduction in metabolic activity in hFOB
272
cells treated with ACP2 siRNA. This result carried through to the ALP assay, in which staining
273
intensity and ALP+ area fraction were dramatically reduced in only the TNAP siRNA and ACP2
274
siRNA groups. The combination of these results with the gene expression data suggests that
275
ACP2 regulates early osteoblast proliferation/viability, but does not directly regulate osteoblast
276
differentiation.
277
We provide evidence that our approach can steer researchers from GWAS signals toward
278
genes relevant to the pathogenesis of the given trait. Furthermore, because our method treats all
279
genes in implicated TADs equally, functional classification extends to the identification of single
280
variant pleiotropic events; as was the case with an intronic FTO variant impacting both IRX3 and
281
IRX5 [8].
282
Despite the advantages presented by TAD_Pathways, the method has a number of
283
limitations. Currently, our method will not overcome the possibility of a gene being
284
inappropriately included in a pathway that it does not actually contribute to, plus all other
285
propagated errors related to pathway curation and analyses [46]. Network based methods built on
286
gene-gene interaction data also suffer from similar biases [47], but potentially to a lesser extent
287
than curated pathways. We include both curated and computationally predicted GO annotations
288
to ameliorate this bias. The computational predictions provide additional support that these genes
289
may be important disease associated genes that we would have missed using only experimentally
13
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
290
validated pathway genes. We are also unable to implicate a gene to a trait if it is not assigned to a
291
curated or predicted pathway, or if it does not fall within a TAD corresponding to a GWAS
292
signal. It is also likely that our approach will not work well with every GWAS. Indeed we are
293
implicating causality to given genes - we are not making a direct connection between the gene
294
and the given variant. Furthermore, our method does not include the possibility of finding genes
295
associated with a disease that is impacted by alternative looping, which has been observed to
296
occur in cancer [48,49] and sickle cell anemia [50]. As research on 3D genome organization
297
increases, it is likely that more diseases will include chromosome looping deficiencies as part of
298
their etiology. Additionally, we used TAD boundaries defined by Dixon et al. 2012. A more
299
recent Hi-C analysis at increased resolution substantially reduced the estimated average size of
300
TADs [51]. Nevertheless, there remains disagreement about how TADs are defined [52]. Despite
301
our method using larger TAD boundaries, thus promoting the inclusion of more presumably false
302
positive genes, we retain the ability to identify biologically logical pathways. The larger
303
boundaries permit us to screen a larger number of candidate genes but makes the method
304
analytically conservative by increasing the pathway overrepresentation signal required to surpass
305
the adjusted significance threshold.
306
The validation screen is also limited: it was performed in a simplified in vitro cell culture
307
system lacking organismal complexity, and the cell line selected is largely tetraploid which may
308
partially compensate for gene knockdown. This is particularly true for the lack of reduction in
309
TNAP gene expression in the TNAP siRNA group, in light of the historical selection of the hFOB
310
cell line based on robust ALP staining in culture [53]. As well, while the TAD approach
311
identifies potentially several GWAS associated genes, herein we only examined two genes per
312
TAD – one immediately adjacent to the GWAS SNP and another that we postulated could play a
14
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
313
role in the skeleton. Further work would need to systematically examine the relative importance
314
of each gene in a TAD.
315
Other recent mechanisms and algorithms used to assign causality from association signals or
316
enhancers to genes typically leverage multiple data types including expression [54] or epigenetic
317
features [24]. For example, TargetFinder uses several high throughput genomic marks to identify
318
features predictive of a chromosome physically looping together enhancers and promoters [24].
319
Looping occurs at sub-TAD level resolutions [55] and sub-TADs are variable across cell types.
320
Therefore, in order for a chromosome looping signature to generalize to GWAS signals, a user
321
must assay tissue-specific and high resolution Hi-C to identify more specific interactions.
322
Alternatively, one could also query variants that affect gene expression in high-throughput and
323
systematically match signals to gene expression [56]. A major limitation to these approaches is
324
that several diseases do not yet have a known tissue source or involve multiple tissues. This is
325
particularly true for osteoporosis whereby multiple cell types, as well as systemic factors,
326
influence bone mass [57]. Therefore, identifying these specific signals may require the
327
procedures to be repeated across competing tissue types. In contrast, a TAD_Pathways analysis
328
is computationally cheap and uses publicly available TAD boundaries and GO terms as a guide
329
for assigning genes to GWAS signals. The method is effective in a wide variety of settings and
330
across tissues because TADs are consistent across cell types [25]. In summary, TAD_Pathways
331
can be used to guide researchers toward the most likely causal gene implicated by a GWAS
332
signal. We have also identified ACP2 as a gene involved in BMD determination, which warrants
333
further investigation.
334
335
Conclusions:
15
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
336
TADs offer a novel tool in the investigation of genome function. We present an approach,
337
called TAD_Pathways, to leverage 3D genomics to prioritize and predict causal genes implicated
338
by GWAS signal. At the foundation of our method is the principle that genomic regions within
339
the same TAD more often interact with each other, and therefore, provide the genomic
340
scaffolding that can impact gene function and gene regulation within each TAD. We applied our
341
method to established BMD GWAS signals. By selecting two GWAS signals and two classes of
342
genes for each signal (nearest gene and predicted gene by TAD_Pathways), we demonstrated
343
that our approach can causally implicate genes kilobases away from their associated GWAS
344
signal. We validated ACP2 (TAD prediction), but not ARHGAP1 (nearest gene), and show that
345
ACP2 influences the proliferation and differentiation of osteoblast cells. We were unable to
346
validate either DEAF1 or BET1L and conclude that neither impacts osteoblast gene expression
347
nor metabolic activity. Whether these genes influence other aspects of skeletal biology cannot be
348
determined within the scope of the current study. Future studies focused on BMD GWAS would
349
explore both osteoblast and osteoclast associated changes. In conclusion, as more information
350
and data is collected regarding 3D genome principles, we propose that algorithms that leverage
351
dynamic 3D structure rather than static linear organization will more accurately predict and
352
discover the basic genomic biology of diseases.
353
354
Methods:
355
Data Integration
356
We used previously identified TAD boundaries for hESC, IMR90, mESC, and mcortex cells
357
for all TAD based analyses [22,58]. To describe the genomic content of TADs, we extracted
358
common SNPs (major allele frequency ≤ 0.95) from the 1000 Genomes Phase III data (2 May
16
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
359
2013 release) [59] and downloaded hg19 Gencode genes [60] and hg19 RepeatMasker repeat
360
elements [61]. We downloaded hg19 FASTA files for all chromosomes as provided by the
361
Genome Reference Consortium [62]. Furthermore, we downloaded the NHGRI-EBI GWAS
362
catalog on 25 February 2016, which holds the significant findings of several GWAS’ for over
363
300 traits [1]. Since the GWAS catalog reports hg38 coordinates, we used the hg38 to hg19
364
UCSC chain file [63] and PyLiftover [64] to convert genome build coordinates to hg19. We
365
assessed relevant expression quantitative trait loci (eQTLs) using all tissues in the NCBI GTEx
366
eQTL Browser [65].
367
368
TAD_Pathways
369
Our TAD_Pathways method is a light-weight approach that uses TAD boundary regions,
370
rather than distance explicitly, to identify putative causal genes. We first build a comprehensive
371
TAD based gene list that consists of all genes that fall inside TADs that are implicated with a
372
GWAS signal (see Figure 2). This gene list assumes that all genes within each signal TAD have
373
an equal likelihood of functional impact on the trait or disease of interest. We then input the
374
TAD based gene list into a WebGestalt overrepresentation analysis [66]. WebGestalt is a
375
webapp that facilitates a pathway analysis interface allowing for quick and custom gene set
376
based analyses. We perform a pathway overrepresentation test for the input TAD based genes
377
against GO biological process, molecular function, and cellular component terms with a
378
background of the human genome. Specifically, this tests if the input gene set is associated with
379
any particular GO term at a higher probability than by chance compared to background genes.
380
We include both experimentally validated and computationally inferred genes in each GO term,
381
which allows the method to discover associations for genes that lack literature support. We
17
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
382
consider genes that are annotated to the most significantly enriched GO term to be the associated
383
set [65].
384
385
386
Cell culture and siRNA transfection
A human fetal osteoblast cell line (ATCC hFOB 1.19 CRL-11372) was obtained and
387
subcultured twice at 34°C, 5% CO2 and 95% relative humidity in 1:1 DMEM/F12 with 2.5mM
388
L-glutamine without phenol red (Gibco 21041025) supplemented with 10% FBS (Atlas USDA
389
F0500D) and 0.3mg/mL G418 sulfate (Gibco 10131035). All experiments were conducted in
390
three temporally separated independent technical replicates from cryopreserved P2 aliquots of
391
these cells. 48 hours prior to transfection, media was switched to a G418-free formulation.
392
Transfections were conducted in single-cell suspension using a commercial siRNA reagent
393
system (Santa Cruz sc-45064) according to the manufacturer’s instructions, with 6μL of siRNA
394
duplex and 4μL transfection reagent in 750μL of transfection media per 100,000 cells. Following
395
trypsinization, cells were counted and divided into one of 8 experimental groups: 1) untreated
396
control, 2) siRNA-negative transfection reagent control, 3) scrambled control siRNA (sc-37007),
397
4) TNAP siRNA (sc-38921), 5) ARHGAP1 siRNA (sc-96477), 6) ACP2 siRNA (sc-96327), 7)
398
BET1L siRNA (sc-97007) or 8) Suppressin (DEAF1) siRNA (sc-76613). Samples for RNA
399
isolation were generated by plating 200,000 cells per well in tissue culture treated 6-well plates
400
(Falcon 353046). Samples for MTT assay and alkaline phosphatase (ALP) staining were
401
generated by plating 50,000 cells per well in tissue culture treated 24-well plates (Falcon
402
353047). Cells were then switched to a 37°C incubator and transfected for 6 hours, at which
403
point transfection cocktails were diluted with 2x hFOB media concentrate. Media was
404
completely changed to 1x standard hFOB media 16 hours later.
18
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
405
406
407
Quantitative PCR
RNA was harvested at Day 4 using TRIzol (Ambion) followed by acid guanidinium
408
thiocyanate-phenol-chloroform extraction and RNeasy (Qiagen) spin-column purification with
409
DNase (Qiagen) then reverse-transcribed using a high-capacity RNA-to-cDNA kit (Applied
410
Biosystems). Duplicate qPCR reactions were conducted on 20ng of whole-RNA template using
411
SYBR Select master mix (Applied Biosystems) in an Applied Biosystems 7500 Fast Real-Time
412
PCR System. Primers for GAPDH, OSX, OCN and IBSP were adopted from the literature and
413
synthesized by R&D Systems. New primer sets spanning exon-exon junctions were designed for
414
ARHGAP1, ACP2, BET1L and DEAF1 using NCBI Primer Blast, verified by melt curve analysis
415
and agarose gel electrophoresis. Sequences follow: ARHGAP1 F-
416
GCGGAAATGGTTGGGGATAG R-CCTTAAGAGAAACCGCGCTC (127bp), ACP2 F-
417
AGCGGGTTCCAGCTTGTTT R-TGGCGGTACAGCAAGGTAAC (165bp), BET1L F-
418
GGATGGCATGGACTCGGATT R-TCCTCTGGAGCCCAAAACAC (254bp), DEAF1 F-
419
GGAAGGAGCAGTCCTGCGTT R-TCACCTTCTCCATCACGCTTT (195bp). Results were
420
analyzed using the 2-ddCt method using GAPDH as a housekeeping gene and reported as (mean ±
421
standard deviation) fold-change versus untreated controls. Statistical significance was
422
determined using 2-way homoscedastic Student’s t-tests versus the scrambled siRNA control,
423
annotated using *p≤0.05 and #p≤0.10. The acronym N.S. stands for “not significant”.
424
425
426
427
MTT metabolic assay
Cellular metabolism/proliferation was assessed using a commercial cell growth
determination kit (Sigma CGD1). At 1 or 4 days post-transfection, media in assigned 24-well
19
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
428
plates was switched to 450μL hFOB media plus 50μL MTT solution. Cells were incubated at
429
37°C for 3.5 hours, after which the media was aspirated and the resulting formazan crystals were
430
solubilized in MTT solvent. Plates were shaken and read in a BioTek Synergy H1 microplate
431
reader. Results are reported as the difference in mean absorbance at 570nm-690nm from Day 1
432
to Day 4. Error bars represent root-mean-square standard deviation from the measurements at
433
both days. Statistical significance was determined using 2-way homoscedastic Student’s t-tests
434
versus the scrambled siRNA control, annotated using *p≤0.05 and #p≤0.10.
435
436
437
Alkaline phosphatase (ALP) staining
Plates were stained at 4 days post-transfection using a commercial ALP kit (Sigma 86C).
438
Dried plates were then imaged at 600dpi using an Epson V370 photo scanner and staining was
439
quantified using mean whole-well intensity measurements in ImageJ using raw output files. ALP
440
area fraction was calculated using color thresholding. Values are reported as mean ± standard
441
deviation. Statistical significance was determined using 2-way homoscedastic Student’s t-tests
442
versus the scrambled siRNA control, annotated using *p≤0.05 and #p≤0.10. Images of plates
443
presented as figures were edited using a warming filter and for brightness/contrast using Adobe
444
Photoshop CS6.
445
446
447
Computational Reproducibility
We provide all of our source code under a permissive open source license and encourage
448
others to modify and build upon our work [67]. Additionally, we provide an accompanying
449
docker image [68] to replicate our computational environment
450
(https://hub.docker.com/r/gregway/tad_pathways/) [69].
20
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
451
452
Declarations:
453
Ethics approval and consent to participate
454
Not applicable.
455
Consent for publication
456
Not applicable.
457
Availability of data and material
458
All data used to construct the TAD_Pathways approach are publically available datasets. We
459
make all software used to develop this approach publically available in a GitHub repository
460
(http://github.com/greenelab/tad_pathways). We also provide a docker image
461
(https://hub.docker.com/r/gregway/tad_pathways/) and archive the GitHub software on
462
Zenodo (https://zenodo.org/record/163950).
463
464
465
Competing interests
The authors declare no competing interests.
Funding
466
This work was supported by the Genomics and Computational Biology Graduate program at
467
The University of Pennsylvania (to G.P.W.); the Gordon and Betty Moore Foundation’s Data
468
Driven Discovery Initiative (grant number GBMF 4552 to C.S.G); the National Institute of
469
Dental & Craniofacial Research (grant number NIH F32DE026346 to D.W.Y.); S.F.A.G is
470
supported by the Daniel B. Burke Endowed Chair for Diabetes Research.
471
Authors' contributions
472
GPW wrote the software, analyzed the data, developed the method and wrote the manuscript;
473
DWY performed the experimental validation and wrote the manuscript; KDH performed the
21
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
474
experimental validation and wrote the manuscript; CSG analyzed the data, developed the
475
method and wrote the manuscript; SFAG analyzed the data, developed the method and wrote
476
the manuscript.
477
Acknowledgements
478
Hannah E. Sexton and Troy L. Mitchell assisted in optimizing siRNA transfection
479
conditions. Daniel Himmelstein and Amy Campbell performed analytical code review.
480
Authors' information (optional)
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
22
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
497
Supplementary Figures:
498
499
500
501
502
503
Supplementary Figure S1: Distributions of genes across topologically associating domains of
four cell types. Bars on the top and right side of the plots represent histograms of each respective
metric. We independently discretized TADs into 50 bins where the distribution of elements is
linear from 5’ to 3’ (bin 0 is the 5’ most end of all TADs).
504
505
506
507
508
509
Supplementary Figure S2: Distributions of all repeat elements across topologically associating
domains of four cell types. Bars on the top and right side of the plots represent histograms of
each respective metric. We independently discretized TADs into 50 bins where the distribution
of elements is linear from 5’ to 3’ (bin 0 is the 5’ most end of all TADs).
23
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
510
511
512
513
514
515
516
517
518
519
520
521
522
Supplementary Figure S3: Distributions of SINE/Alu elements and potential confounding
factors across topologically associating domains of four cell types. We independently discretized
TADs into 50 bins where the distribution of elements is linear from 5’ to 3’ (bin 0 is the 5’ most
end of all TADs). The distribution of Alu elements is strongly clustered toward TAD boundaries
but neither the percentage of GC content across TADs nor the evolutionary divergence of Alu
elements can explain the Alu distribution.
Supplementary Figure S4: Bone Mineral Density gene counts associated with various classes
of evidence. The gene LRP5 is implicated by all evidence codes. About 1/3 of all TAD_Pathways
implicated genes are the nearest gene implicated by the GWAS signal.
24
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
523
Tables:
524
Supplementary Table S1: Chi square testing the enrichment of genes at TAD boundaries and
525
enrichment of SNPs towards the right half of TADs.
526
Enrichment of SNPs in the Right of TADs
Left
Right
Chi-square
P
hESC
3268781 3296450
116.6
3.5E-27
IMR90
3211140 3210251
0.12
0.73
mESC*
25161934 25129279
21.2
4.1E-06
mcortex
24616532 24641931
13.1
3.0E-4
* Note- mouse embryonic stem cells have an opposite enrichment. There are more SNPs in the
527
left of TADs
528
529
We provide Supplementary Tables S2 and S3 as attached .xls files.
530
531
Abbreviations:
532
TAD – Topologically associating domain
533
GWAS – Genome wide association study
534
SNP – Single nucleotide polymorphism
535
BMD – Bone mineral density
536
hESC – Human embryonic stem cells
537
mESC – Mouse embryonic stem cells
538
IMR90 – Human fibroblast cells
539
mcortex – Mouse cortex cells
540
siRNA – Small interfering RNA
541
eQTL – Expression quantitative trail loci
542
hFOB – Human fetal osteoblast
25
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
543
TNAP - tissue-nonspecific alkaline phosphatase
544
OCN – osteocalcin
545
IBSP – bone sialoprotein
546
ALP – Alkaline phosphatase
547
GO – Gene ontology
548
eQTL – Expression quantitative trait loci
549
550
References:
551
552
1. Welter D, MacArthur J, Morales J, Burdett T, Hall P, Junkins H, et al. The NHGRI GWAS
Catalog, a curated resource of SNP-trait associations. Nucleic Acids Res. 2014;42:D1001–6.
553
554
2. Zhang F, Lupski JR. Non-coding genetic variants in human disease. Hum. Mol. Genet.
2015;24:R102–10.
555
556
3. Ward LD, Kellis M. Interpreting noncoding genetic variation in complex traits and human
disease. Nat. Biotechnol. 2012;30:1095–106.
557
558
559
4. McVean GA, Altshuler (Co-Chair) DM, Durbin (Co-Chair) RM, Abecasis GR, Bentley DR,
Chakravarti A, et al. An integrated map of genetic variation from 1,092 human genomes. Nature.
2012;491:56–65.
560
561
5. Brodie A, Azaria JR, Ofran Y. How far from the SNP may the causative genes be? Nucleic
Acids Res. 2016;gkw500.
562
563
6. Tung YCL, Yeo GSH, O’Rahilly S, Coll AP. Obesity and FTO: Changing Focus at a Complex
Locus. Cell Metab. 2014;20:710–8.
564
565
566
7. Ragvin A, Moro E, Fredman D, Navratilova P, Drivenes O, Engstrom PG, et al. Long-range
gene regulation links genomic type 2 diabetes and obesity risk regions to HHEX, SOX4, and
IRX3. Proc. Natl. Acad. Sci. 2010;107:775–80.
567
568
8. Claussnitzer M, Dankel SN, Kim K-H, Quon G, Meuleman W, Haugen C, et al. FTO Obesity
Variant Circuitry and Adipocyte Browning in Humans. N. Engl. J. Med. 2015;373:895–907.
569
570
571
9. Smemo S, Tena JJ, Kim K-H, Gamazon ER, Sakabe NJ, Gómez-Marín C, et al. Obesityassociated variants within FTO form long-range functional connections with IRX3. Nature.
2014;507:371–5.
26
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
572
573
574
10. Grant SFA, Thorleifsson G, Reynisdottir I, Benediktsson R, Manolescu A, Sainz J, et al.
Variant of transcription factor 7-like 2 (TCF7L2) gene confers risk of type 2 diabetes. Nat.
Genet. 2006;38:320–3.
575
576
577
11. Xia Q, Chesi A, Manduchi E, Johnston BT, Lu S, Leonard ME, et al. The type 2 diabetes
presumed causal variant within TCF7L2 resides in an element that controls the expression of
ACSL5. Diabetologia. 2016;59:2360–8.
578
579
12. Herman MA, Rosen ED. Making Biological Sense of GWAS Data: Lessons from the FTO
Locus. Cell Metab. 2015;22:538–9.
580
581
13. Dekker J, Rippe K, Dekker M, Kleckner N. Capturing chromosome conformation. Science.
2002;295:1306–11.
582
14. Dekker J. Gene Regulation in the Third Dimension. Science. 2008;319:1793–4.
583
584
585
15. Lieberman-Aiden E, van Berkum NL, Williams L, Imakaev M, Ragoczy T, Telling A, et al.
Comprehensive Mapping of Long-Range Interactions Reveals Folding Principles of the Human
Genome. Science. 2009;326:289–93.
586
587
16. Duan Z, Andronescu M, Schutz K, McIlwain S, Kim YJ, Lee C, et al. A three-dimensional
model of the yeast genome. Nature. 2010;465:363–7.
588
589
590
17. Phillips-Cremins JE, Sauria MEG, Sanyal A, Gerasimova TI, Lajoie BR, Bell JSK, et al.
Architectural Protein Subclasses Shape 3D Organization of Genomes during Lineage
Commitment. Cell. 2013;153:1281–95.
591
592
593
18. Sexton T, Yaffe E, Kenigsberg E, Bantignies F, Leblanc B, Hoichman M, et al. ThreeDimensional Folding and Functional Organization Principles of the Drosophila Genome. Cell.
2012;148:458–72.
594
595
19. Symmons O, Uslu VV, Tsujimura T, Ruf S, Nassari S, Schwarzer W, et al. Functional and
topological characteristics of mammalian regulatory domains. Genome Res. 2014;24:390–400.
596
597
20. Pope BD, Ryba T, Dileep V, Yue F, Wu W, Denas O, et al. Topologically associating
domains are stable units of replication-timing regulation. Nature. 2014;515:402–5.
598
599
600
21. Gurudatta B, Yang J, Van Bortle K, Donlin-Asp P, Corces V. Dynamic changes in the
genomic localization of DNA replication-related element binding factor during the cell cycle.
Cell Cycle. 2013;12:1605–15.
601
602
22. Dixon JR, Selvaraj S, Yue F, Kim A, Li Y, Shen Y, et al. Topological domains in
mammalian genomes identified by analysis of chromatin interactions. Nature. 2012;485:376–80.
603
604
605
23. Nora EP, Dekker J, Heard E. Segmental folding of chromosomes: A basis for structural and
regulatory chromosomal neighborhoods?: Prospects &amp; Overviews. BioEssays.
2013;35:818–28.
27
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
606
607
24. Whalen S, Truty RM, Pollard KS. Enhancer–promoter interactions are encoded by complex
genomic signatures on looping chromatin. Nat. Genet. 2016;48:488–96.
608
609
610
25. Smith EM, Lajoie BR, Jain G, Dekker J. Invariant TAD Boundaries Constrain Cell-TypeSpecific Looping Interactions between Promoters and Distal Elements around the CFTR Locus.
Am. J. Hum. Genet. 2016;98:185–201.
611
612
613
26. Richards JB, Rivadeneira F, Inouye M, Pastinen TM, Soranzo N, Wilson SG, et al. Bone
mineral density, osteoporosis, and osteoporotic fractures: a genome-wide association study.
Lancet Lond. Engl. 2008;371:1505–12.
614
615
616
27. Rivadeneira F, Styrkársdottir U, Estrada K, Halldórsson BV, Hsu Y-H, Richards JB, et al.
Twenty bone-mineral-density loci identified by large-scale meta-analysis of genome-wide
association studies. Nat. Genet. 2009;41:1199–206.
617
618
619
28. Estrada K, Styrkarsdottir U, Evangelou E, Hsu Y-H, Duncan EL, Ntzani EE, et al. Genomewide meta-analysis identifies 56 bone mineral density loci and reveals 14 loci associated with
risk of fracture. Nat. Genet. 2012;44:491–501.
620
621
622
29. Styrkarsdottir U, Thorleifsson G, Sulem P, Gudbjartsson DF, Sigurdsson A, Jonasdottir A, et
al. Nonsense mutation in the LGR4 gene is associated with several human diseases and other
traits. Nature. 2013;497:517–20.
623
624
30. Hendrickx G, Boudin E, Van Hul W. A look behind the scenes: the risk and pathogenesis of
primary osteoporosis. Nat. Rev. Rheumatol. 2015;11:462–74.
625
626
31. Choo SY. The HLA System: Genetics, Immunology, Clinical Testing, and Clinical
Implications. Yonsei Med. J. 2007;48:11.
627
628
629
32. Håvik B, Le Hellard S, Rietschel M, Lybæk H, Djurovic S, Mattheisen M, et al. The
Complement Control-Related Genes CSMD1 and CSMD2 Associate to Schizophrenia. Biol.
Psychiatry. 2011;70:35–42.
630
631
33. Kwon E, Wang W, Tsai L-H. Validation of schizophrenia-associated genes CSMD1,
C10orf26, CACNA1C and TCF4 as miR-137 targets. Mol. Psychiatry. 2013;18:11–2.
632
633
34. Wang K, Li M, Bucan M. Pathway-Based Approaches for Analysis of Genomewide
Association Studies. Am. J. Hum. Genet. 2007;81:1278–83.
634
635
35. Hao K, Di X, Cawley S. LdCompare: rapid computation of single- and multiple-marker r2
and genetic coverage. Bioinformatics. 2007;23:252–4.
636
637
638
36. Taşan M, Musso G, Hao T, Vidal M, MacRae CA, Roth FP. Selecting causal genes from
genome-wide association studies via functionally coherent subnetworks. Nat. Methods.
2014;12:154–9.
639
640
37. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, et al. Gene Ontology:
tool for the unification of biology. Nat. Genet. 2000;25:25–9.
28
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
641
642
643
38. Petersen A, Alvarez C, DeClaire S, Tintle NL. Assessing Methods for Assigning SNPs to
Genes in Gene-Based Tests of Association Using Common Variants. Chen L, editor. PLoS ONE.
2013;8:e62161.
644
645
646
39. Mu XJ, Lu ZJ, Kong Y, Lam HYK, Gerstein MB. Analysis of genomic variation in noncoding elements using population-scale sequencing data from the 1000 Genomes Project.
Nucleic Acids Res. 2011;39:7058–76.
647
648
40. Jurka J, Kohany O, Pavlicek A, Kapitonov VV, Jurka MV. Duplication, coclustering, and
selection of human Alu retrotransposons. Proc. Natl. Acad. Sci. 2004;101:1268–72.
649
41. Deininger P. Alu elements: know the SINEs. Genome Biol. 2011;12:236.
650
651
652
42. Wang J, Vicente-García C, Seruggia D, Moltó E, Fernandez-Miñán A, Neto A, et al. MIR
retrotransposon sequences provide insulators to the human genome. Proc. Natl. Acad. Sci.
2015;112:E4428–37.
653
654
655
43. Gu Z, Jin K, Crabbe MJC, Zhang Y, Liu X, Huang Y, et al. Enrichment analysis of Alu
elements with different spatial chromatin proximity in the human genome. Protein Cell.
2016;7:250–66.
656
657
44. Baron R, Kneissel M. WNT signaling in bone homeostasis and disease: from human
mutations to treatments. Nat. Med. 2013;19:179–92.
658
659
45. Krishnan V, Bryant HU, Macdougald OA. Regulation of bone mass by Wnt signaling. J.
Clin. Invest. 2006;116:1202–9.
660
661
46. Khatri P, Sirota M, Butte AJ. Ten Years of Pathway Analysis: Current Approaches and
Outstanding Challenges. Ouzounis CA, editor. PLoS Comput. Biol. 2012;8:e1002375.
662
663
664
47. Reguly T, Breitkreutz A, Boucher L, Breitkreutz B-J, Hon GC, Myers CL, et al.
Comprehensive curation and analysis of global interaction networks in Saccharomyces
cerevisiae. J. Biol. 2006;5:11.
665
666
48. Corces MR, Corces VG. The three-dimensional cancer genome. Curr. Opin. Genet. Dev.
2016;36:1–7.
667
668
669
49. Flavahan WA, Drier Y, Liau BB, Gillespie SM, Venteicher AS, Stemmer-Rachamimov AO,
et al. Insulator dysfunction and oncogene activation in IDH mutant gliomas. Nature.
2015;529:110–4.
670
671
672
50. Deng W, Lee J, Wang H, Miller J, Reik A, Gregory PD, et al. Controlling long-range
genomic interactions at a native locus by targeted tethering of a looping factor. Cell.
2012;149:1233–44.
673
674
675
51. Rao SSP, Huntley MH, Durand NC, Stamenova EK, Bochkov ID, Robinson JT, et al. A 3D
Map of the Human Genome at Kilobase Resolution Reveals Principles of Chromatin Looping.
Cell. 2014;159:1665–80.
29
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
676
677
52. Bonev B, Cavalli G. Organization and function of the 3D genome. Nat. Rev. Genet.
2016;17:661–78.
678
679
680
53. Harris SA, Enger RJ, Riggs BL, Spelsberg TC. Development and characterization of a
conditionally immortalized human fetal osteoblastic cell line. J. Bone Miner. Res. Off. J. Am.
Soc. Bone Miner. Res. 1995;10:178–86.
681
682
683
54. Zhu Z, Zhang F, Hu H, Bakshi A, Robinson MR, Powell JE, et al. Integration of summary
data from GWAS and eQTL studies predicts complex trait gene targets. Nat. Genet.
2016;48:481–7.
684
685
686
55. Dowen JM, Fan ZP, Hnisz D, Ren G, Abraham BJ, Zhang LN, et al. Control of Cell Identity
Genes Occurs in Insulated Neighborhoods in Mammalian Chromosomes. Cell. 2014;159:374–
87.
687
688
689
56. Tewhey R, Kotliar D, Park DS, Liu B, Winnicki S, Reilly SK, et al. Direct Identification of
Hundreds of Expression-Modulating Variants using a Multiplexed Reporter Assay. Cell.
2016;165:1519–29.
690
691
57. Lupsa BC, Insogna K. Bone Health and Osteoporosis. Endocrinol. Metab. Clin. North Am.
2015;44:517–30.
692
693
58. Ho JWK, Jung YL, Liu T, Alver BH, Lee S, Ikegami K, et al. Comparative analysis of
metazoan chromatin organization. Nature. 2014;512:449–52.
694
695
59. 1000 Genomes Project Consortium, Auton A, Brooks LD, Durbin RM, Garrison EP, Kang
HM, et al. A global reference for human genetic variation. Nature. 2015;526:68–74.
696
697
698
60. Harrow J, Frankish A, Gonzalez JM, Tapanari E, Diekhans M, Kokocinski F, et al.
GENCODE: The reference human genome annotation for The ENCODE Project. Genome Res.
2012;22:1760–74.
699
700
61. Smit A, Hubley R, Green P. RepeatMasker Open-4.0 [Internet]. Available from:
http://www.repeatmasker.org
701
702
62. Church DM, Schneider VA, Graves T, Auger K, Cunningham F, Bouk N, et al. Modernizing
Reference Genome Assemblies. PLoS Biol. 2011;9:e1001091.
703
704
63. Kent WJ, Sugnet CW, Furey TS, Roskin KM, Pringle TH, Zahler AM, et al. The Human
Genome Browser at UCSC. Genome Res. 2002;12:996–1006.
705
706
64. Tretyakov K. pyliftover [Internet]. 2014. Available from:
https://github.com/konstantint/pyliftover
707
708
65. GTEx Consortium. The Genotype-Tissue Expression (GTEx) project. Nat. Genet.
2013;45:580–5.
30
bioRxiv preprint first posted online Nov. 15, 2016; doi: http://dx.doi.org/10.1101/087718. The copyright holder for this preprint (which was not
peer-reviewed) is the author/funder. It is made available under a CC-BY 4.0 International license.
709
710
66. Wang J, Duncan D, Shi Z, Zhang B. WEB-based GEne SeT AnaLysis Toolkit (WebGestalt):
update 2013. Nucleic Acids Res. 2013;41:W77–83.
711
712
67. Gregory Way, Casey Green. Greenelab/Tad_Pathways: Pre-Release. 2016 [cited 2016 Nov
14]; Available from: https://doi.org/10.5281/zenodo.163950
713
714
68. Boettiger C. An introduction to Docker for reproducible research. ACM SIGOPS Oper. Syst.
Rev. 2015;49:71–9.
715
716
69. Gregory Way, Struan Grant, Casey Greene. TAD Pathways Archived Docker Image. 2016
[cited 2016 Nov 14]; Available from: https://doi.org/10.5281/zenodo.166556
717
31